Auditing the Judge: Internal Signals for LLM Reasoning Assurance
Organizations: Freie Universität Berlin
Abstract
Language-model systems use self-judgment to decide which answers to release, review, or recompute. Because intervention is costly, an assurance score must rank errors well at the available budget, not merely over the full ranking. We compare the judge's continuous verdict-token score with a supervised correctness probe trained on attention-head activations from the same prompt. Exact identities link fixed-budget error recall to residual risk and show that AUROC averages recall over all budgets. Across three open 7-12B models on GSM8K and MATH, the probe finds 1.3-22.5 additional percentage points of errors at 5% review in all six model-dataset settings (median 7.4 points). At 10% routing, probe-guided allocation improves accuracy over equal-cost random allocation in both GSM8K experiments, but beats judge-guided routing only for the weaker judge. Cross-dataset transfer lowers probe AUROC for every model in both directions relative to target-domain training. Because dataset, weight precision, and generation budget change together, this measures a joint protocol shift rather than a dataset-only effect. Internal signals can improve supervised, in-domain budget allocation, but are not domain-general confidence estimators.
Figures & tables
| AUROC (correctness) | AP (error class) | ||||||||
| Model | Data | probe | judge | [95% CI] | probe | judge | [95% CI] | ||
| Gemma 3 12B | GSM8K | 1319 | 0.061 | 0.879 | 0.848 | +0.030 [+0.001, +0.063] | 0.498 | 0.295 | +0.203 [+0.101, +0.292] |
| Qwen2.5 7B | GSM8K | 1319 | 0.081 | 0.858 | 0.873 | -0.015 [-0.048, +0.017] | 0.538 | 0.419 | +0.120 [+0.028, +0.201] |
| Llama 3.1 8B | GSM8K | 1307 | 0.178 | 0.813 | 0.697 | +0.116 [+0.079, +0.157] | 0.584 | 0.350 | +0.233 [+0.169, +0.296] |
| Gemma 3 12B | MATH | 4883 | 0.184 | 0.914 | 0.822 | +0.092 [+0.077, +0.107] | 0.776 | 0.539 | +0.237 [+0.205, +0.267] |
| Qwen2.5 7B | MATH | 4949 | 0.270 | 0.888 | 0.848 | +0.040 [+0.030, +0.050] | 0.771 | 0.690 | +0.081 [+0.061, +0.103] |
| Error recall at 5% review | Probe advantage over judge | ||||
|---|---|---|---|---|---|
| Model | Data | probe | judge | recall | unreviewed error-rate reduction (pp) [95% CI] |
| Gemma 3 12B | GSM8K | 0.487 | 0.263 | +0.225 | +1.44 [+0.80, +1.92] |
| Qwen2.5 7B | GSM8K | 0.402 | 0.308 | +0.093 | +0.80 [+0.16, +1.52] |
| Llama 3.1 8B | GSM8K | 0.227 | 0.120 | +0.107 | +2.01 [+1.21, +2.66] |
| Gemma 3 12B | MATH | 0.254 | 0.200 | +0.055 | +1.06 [+0.80, +1.38] |
| Qwen2.5 7B | MATH | 0.170 | 0.157 | +0.013 | +0.38 [+0.13, +0.64] |
| deployment comparison | primary-probe transfer | transferred ablations | ||||||
|---|---|---|---|---|---|---|---|---|
| Model | Target | judge | probe | vs. judge | in-domain | vs. in-dom. | controlled | shallow |
| Gemma 3 12B | GSM8K | 0.848 | 0.808 | -0.040 | 0.879 | -0.071 | 0.802 | 0.738 |
| MATH | 0.822 | 0.808 | -0.014 | 0.914 | -0.106 | 0.567 | 0.801 | |
| Qwen2.5 7B | GSM8K | 0.873 | 0.819 | -0.054 | 0.858 | -0.040 | 0.774 | 0.711 |
| MATH | 0.848 | 0.763 | -0.085 | 0.888 | -0.125 | 0.624 | 0.772 | |
| Llama 3.1 8B | GSM8K | 0.697 | 0.757 | +0.060 | 0.813 | -0.055 | 0.723 | 0.692 |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Identity | Result |
|---|---|
| Residual risk | Agreement to floating-point precision in all 42 cells. |
| AUROC | Difference equals the exact endpoint term to numerical precision. |
| Routing | Oracle curves agree to floating-point precision in both arms. |
| Run | score | LHS | ||
|---|---|---|---|---|
| Gemma 3/GSM8K | probe | 0.8787 | 0.8561 | 0.8557 |
| Gemma 3/GSM8K | judge | 0.8484 | 0.8276 | 0.8272 |
| Qwen2.5/GSM8K | probe | 0.8583 | 0.8296 | 0.8293 |
| Qwen2.5/GSM8K | judge | 0.8731 | 0.8432 | 0.8428 |
| Llama 3.1/GSM8K | probe | 0.8127 | 0.7574 | 0.7570 |
| Llama 3.1/GSM8K | judge | 0.6969 | 0.6622 | 0.6618 |
| Run | split | gen. | trunc | no gold | analysed | ||
| Gemma 3 12B / GSM8K | train | 7473 | 7 | 0 | 7466 | 0.057 | 0.057 |
| test | 1319 | 0 | 0 | 1319 | 0.061 | 0.061 | |
| Qwen2.5 7B / GSM8K | train | 7473 | 0 | 0 | 7473 | 0.050 | 0.050 |
| test | 1319 | 0 | 0 | 1319 | 0.081 | 0.081 | |
| Llama 3.1 8B / GSM8K | train | 7473 | 30 | 0 | 7443 | 0.107 | 0.110 |
| test | 1319 | 12 | 0 | 1307 | 0.178 | 0.185 |
| Model | Target | judge | detector | in | vs. in-dom. [95% CI] | vs. judge [95% CI] | |
|---|---|---|---|---|---|---|---|
| Gemma 3 12B | GSM8K | 0.848 | raw MLP | 0.879 | 0.808 | -0.071 [-0.109, -0.034] † | -0.040 [-0.077, -0.002] † |
| controlled | 0.844 | 0.802 | -0.041 [-0.097, +0.017] | -0.046 [-0.093, -0.002] † | |||
| shallow | 0.756 | 0.738 | -0.018 [-0.048, +0.013] | -0.109 [-0.168, -0.054] † | |||
| MATH | 0.822 | raw MLP | 0.914 | 0.808 | -0.106 [-0.122, -0.091] † | -0.014 [-0.027, -0.001] † | |
| controlled | 0.902 | 0.567 | -0.335 [-0.360, -0.310] † | -0.255 [-0.280, -0.230] † | |||
| shallow | 0.801 | 0.801 | -0.000 [-0.005, +0.005] | -0.021 [-0.035, -0.007] † |
Language Model Error Detection
Language model error detection identifies unreliable answers, reasoning steps, and actions so systems can flag, verify, or repair them.
See all 113 papers